Optimize XLSX subtable detection memory usage

[CyMule]

Summary

Optimizes XLSX subtable detection by replacing dense worksheet-sized NetworkX graph construction with a sparse traversal over populated cells.

The public partition_xlsx() behavior is intended to remain unchanged:

• populated cells are grouped by 4-neighbor connectivity
• existing row-overlap merge behavior is preserved
• emitted element text and table HTML metadata are covered by generated XLSX tests

Also removes networkx from the xlsx runtime extra, since it is now only needed by the local benchmark's dense reference implementation.

Benchmark

Fresh subprocess benchmark on generated .xlsx files:

Case	Components	Sparse Time	Dense Time	Sparse Peak Delta	Dense Peak Delta
dense_table	1 / 1	0.0103s	0.0339s	5.8 MB	19.8 MB
sparse_wide_edges	26 / 26	0.0039s	6.5139s	2.9 MB	1578.1 MB
separated_blocks	23 / 23	0.0032s	0.7332s	0.5 MB	208.8 MB

Testing

• uv run --extra xlsx --group test pytest -q test_unstructured/partition/test_xlsx.py
• uv run ruff check unstructured/partition/xlsx.py test_unstructured/partition/test_xlsx.py scripts/performance/benchmark_xlsx_connected_components.py
• uv run --extra xlsx --with networkx python scripts/performance/benchmark_xlsx_connected_components.py --repeat 3

---

Summary by cubic

Optimized XLSX subtable detection by replacing a dense worksheet-sized graph with a sparse traversal of populated cells, reducing peak memory and speeding up processing. Public partition_xlsx() output is unchanged.

• Refactors

  • Replaced dense networkx grid with a sparse 4-neighbor traversal over non-empty cells; preserved row-overlap merge.
  • Added generated-XLSX tests to verify element text, HTML metadata, and page info.
  • Added a local benchmark utility.

• Dependencies

  • Removed networkx from the xlsx extra; it’s now only used by the local benchmark; updated uv.lock.
  • Bumped version to 0.22.32 and added a CHANGELOG entry.

^{Written for commit 40b5173. Summary will update on new commits.}